Physics-Assisted Long-Short-Term-Memory Network for Forecasting of Fouling in a Regenerative Heat Exchanger

Abstract

Air Preheater (APH) is a regenerative heat exchanger employed in power plants for improving the boiler thermal efficiency. Fouling of APH is a serious problem as it deteriorates heat transfer efficiency and causes unplanned shutdowns. This complex physico-chemical phenomenon is governed by APH operating conditions, flue gas composition and ambient conditions. We propose a physics-assisted Long-Short-Term-Memory (LSTM) network model to forecast the fouling of APH. A physics-based soft sensor, indicative of chemical deposition within the APH, is used as an additional feature. The physics-assisted basic and autoregressive LSTM models are found to be more accurate than the basic and autoregressive LSTM models, owing to additional insights coming from the physics-based soft sensor. They can help in effective predictive maintenance of APH by preempting forced outages of the plant due to fouling, up to three months in advance. The proposed framework can be easily adapted for forecasting of fouling in heat exchangers used in diverse industries.